1. Field
The present invention relates to the field of digital watermarking.
2. General Background
Each day, consumers receive information from content providers in a variety of forms such as movies or television programming. When transmitted in a standard analog format, the information is somewhat protected due to its inherent analog nature. For example, a recorded movie normally has poorer image quality than the original analog version.
Over the last few years, there has been a growing demand for digital content. In general, xe2x80x9cdigital contentxe2x80x9d is data that is perceived during display or playback (e.g., a still image, video, text, audio, programs, etc.). Unlike analog, digital content can be easily copied or manipulated without affecting its perceived quality. This xe2x80x9cqualityxe2x80x9d may be measured through visual clarity during display, audible clarity during audio playback, accuracy of characters in text or code as well as other factors. Since digital content can be easily manipulated, content providers have been hesitant in providing premier digital content due to the absence of any mechanism to protect the content against illegal copying and/or distribution.
Recently, digital watermarking has emerged as a technique to protect against unauthorized copying and distribution of digital content. More specifically, xe2x80x9cdigital watermarkingxe2x80x9d comprises an act of inserting information (referred to as a xe2x80x9cwatermarkxe2x80x9d) into content either in a spatial domain or in a compressed domain. The insertion of a watermark may be performed in an unobtrusive way so that the quality of the content is not reduced. For example, an invisible watermark is inserted into content in the spatial domain by adding or subtracting small numbers from pixel values. When watermarks are inserted into content in the compressed domain, however, values are added or subtracted from transform coefficients.
Digital watermarking further comprises the act of detecting and subsequently extracting a watermark from content, even if the content is altered with significant degradation. Conventional detection methods use a spread spectrum approach as set forth in equation (2). As denoted, xe2x80x9cBjxe2x80x9d is a signal with values of xe2x88x921 or 1, where xe2x80x9cjxe2x80x9d refers to a particular signal bit, xe2x80x9cSixe2x80x9d as a spread sequence of signal bit Bj where xe2x80x9cixe2x80x9d refers to a number (N) of pixels or blocks of pixels used to encode bit Bj. xe2x80x9cAixe2x80x9d is an amplitude factor, and Ri is a pseudo-random number modulation sequence with a series of values being xe2x80x9cxe2x88x921xe2x80x9d or xe2x80x9c1xe2x80x9d. The addition of watermark (Wi) changes the intensity of an image from its normal intensity (Ii) to its watermarked intensity (Ii) as set forth in equation (1):
Ii+Wi=Ii=SiAiRixe2x80x83xe2x80x83(1) 
Thus, as set forth in equation (2), Bj can be detected by multiplying the watermarked pixels Ii by Ri.                               B          j                =                              sign            ⁡                          (                                                ∑                                      i                    =                    1                                    N                                ⁢                                                      I                    i                    xe2x80x2                                    ⁢                                      R                    i                                                              )                                =                      sign            ⁡                          (                                                                    ∑                                          i                      =                      1                                        N                                    ⁢                                                            I                      i                                        ⁢                                          R                      i                                                                      +                                                      ∑                                          i                      =                      1                                        N                                    ⁢                                                            S                      i                                        ⁢                                          A                      i                                        ⁢                                          R                      i                      2                                                                                  )                                                          (        2        )            
Thus, the value of the first term       ∑          i      =      1        N    ⁢            I      i        ⁢          R      i      
should approach zero when xe2x80x9cNxe2x80x9d (the number of encoded pixels) is large so that the sign of Bj is equal to the sign of the second term       ∑          i      =      1        N    ⁢            S      i        ⁢          A      i        ⁢                  R        i        2            .      
Since, in practice,       ∑          i      =      1        N    ⁢            I      i        ⁢          R      i      
is not truly zero, Bj is calculated as follows:                               B          j                =                  sign          ⁡                      (                                                            ∑                                      i                    =                    1                                    N                                ⁢                                                      I                    i                    xe2x80x2                                    ⁢                                      R                    i                                                              -                                                mean                  ⁡                                      (                                          I                      i                      xe2x80x2                                        )                                                  xc3x97                                                      ∑                                          i                      =                      1                                        N                                    ⁢                                      R                    i                                                                        )                                              (        3        )            
These conventional watermarking functions have the desired features of efficient detection, no requirement of original material for detection and little impact on compressibility. However, they do not address invisibility, high detection reliability and robustness. For example, a watermark can be kept invisible by determining an appropriate amplitude factor value. As set forth on pages 92-99 of a 1997 reference by Ingemar J. Cox and Matt L. Miller entitled xe2x80x9cA Review of Watermarks and the Importance of Perceptual Modeling,xe2x80x9d SPIE vol. 3016, one technique for determining the amplitude factor value is based on properties of the image such as its contrast and brightness. However, this technique does not account for watermarks that are invisible in still images but produce artifacts in video.
The conventional watermarking functions have an inherent weakness in that detection reliability is limited where the term       ∑          i      =      1        N    ⁢            I      i        ⁢          R      i      
dominates the term       ∑          i      =      1        N    ⁢            S      i        ⁢          A      i        ⁢                  R        i        2            .      
Normally, each individual IiRi product is about a factor of 100 larger than each SiAiRi2 product. Thus, if N is not sufficiently large so that the IiRi product approaches zero, a substantial percentage of the sign detection for Bj will be incorrect. While proposals have been made to increase N to improve detection reliability, this would drastically increase the number of frames required for reliable detection.
It is now appreciated that conventional watermarking techniques do not fully address invisibility and security considerations. For example, a watermark invisible in a still image may produce visible artifacts in a video sequence when the sign of the watermark changes between adjacent frames. Also, watermarks produced from the same key may be detected when inserted into successive frames of the video sequence by averaging images found in successive frames.
Hence, it would be desirable to develop digital watermarking mechanisms supporting improved detection reliability through frame differences or image-dependent watermark generation.
Briefly, one embodiment of the invention relates to a method for improving detection of a watermark. The method comprises generating a pseudo-random sequence of numbers based on data associated with the data set, producing a watermark based on the pseudo-random number sequence, and embedding the watermark into the data set.